diff options
| author | skal <pascal.massimino@gmail.com> | 2026-03-25 10:05:42 +0100 |
|---|---|---|
| committer | skal <pascal.massimino@gmail.com> | 2026-03-25 10:05:42 +0100 |
| commit | ce6e5b99f26e4e7c69a3cacf360bd0d492de928c (patch) | |
| tree | a8d64b33a7ea1109b6b7e1043ced946cac416756 /cnn_v3/docs/HOWTO.md | |
| parent | 8b4d7a49f038d7e849e6764dcc3abd1e1be01061 (diff) | |
feat(cnn_v3): 3×3 dilated bottleneck + Sobel loss + FiLM warmup + architecture PNG
- Replace 1×1 pointwise bottleneck with Conv(8→8, 3×3, dilation=2):
effective RF grows from ~13px to ~29px at ¼res (~+1 KB weights)
- Add Sobel edge loss in training (--edge-loss-weight, default 0.1)
- Add FiLM 2-phase training: freeze MLP for warmup epochs then
unfreeze at lr×0.1 (--film-warmup-epochs, default 50)
- Update weight layout: BN 72→584 f16, total 1964→2476 f16 (4952 B)
- Cascade offsets in C++ effect, JS tool, export/gen_test_vectors scripts
- Regenerate test_vectors.h (1238 u32); parity max_err=9.77e-04
- Generate dark-theme U-Net+FiLM architecture PNG (gen_architecture_png.py)
- Replace ASCII art in CNN_V3.md and HOW_TO_CNN.md with PNG embed
handoff(Gemini): bottleneck dilation + Sobel loss + FiLM warmup landed.
Next: run first real training pass (see cnn_v3/docs/HOWTO.md §3).
Diffstat (limited to 'cnn_v3/docs/HOWTO.md')
| -rw-r--r-- | cnn_v3/docs/HOWTO.md | 6 |
1 files changed, 3 insertions, 3 deletions
diff --git a/cnn_v3/docs/HOWTO.md b/cnn_v3/docs/HOWTO.md index 1aead68..9a3efdf 100644 --- a/cnn_v3/docs/HOWTO.md +++ b/cnn_v3/docs/HOWTO.md @@ -439,10 +439,10 @@ FiLM γ/β are computed CPU-side by the FiLM MLP (Phase 4) and uploaded each fra |-------|---------|------|-----------| | enc0 | 20×4×9=720 | +4 | 724 | | enc1 | 4×8×9=288 | +8 | 296 | -| bottleneck | 8×8×1=64 | +8 | 72 | +| bottleneck | 8×8×9=576 | +8 | 584 | | dec1 | 16×4×9=576 | +4 | 580 | | dec0 | 8×4×9=288 | +4 | 292 | -| **Total** | | | **2064 f16 = ~4 KB** | +| **Total** | | | **2476 f16 = ~4.84 KB** | **Asset IDs** (registered in `workspaces/main/assets.txt` + `src/effects/shaders.cc`): `SHADER_CNN_V3_COMMON`, `SHADER_CNN_V3_ENC0`, `SHADER_CNN_V3_ENC1`, @@ -725,7 +725,7 @@ Both tools print first 8 pixels in the same format: **Expected delta:** ≤ 1/255 (≈ 4e-3) per channel, matching the parity test (`test_cnn_v3_parity`). Larger deltas indicate a weight mismatch — re-export -with `export_cnn_v3_weights.py` and verify the `.bin` size is 3928 bytes. +with `export_cnn_v3_weights.py` and verify the `.bin` size is 4952 bytes. --- |